HowISwitch 🚀 | From AI conversations to collaborative AI engineering
What if AI agents could work together like an actual engineering team?
That question inspired me to build ShipRoom AI for the WeMakeDevs Flip the Switch Hackathon 2026.
I used Switch to bring specialized AI agents into a shared engineering workflow, where they could investigate defects, coordinate remediation, review evidence, and help assess software release readiness.
Here's how I put Switch to work:
🔍 1. Connected specialized agents
I configured the Codex provider and registered two agents:
-
release-auditor– investigates source code, identifies defects, highlights edge cases, and independently reviews fixes. -
release-preparer– coordinates findings, tracks blockers, and prepares release-readiness reports.
💬 2. Created a shared collaboration space
I set up a private Mattermost channel called shiproom-ai, bringing the agents and me into one place to coordinate investigations and handoffs.
🤝 3. Designed a multi-agent workflow
Rather than asking one AI assistant to do everything, I gave each agent a clear responsibility.
The Auditor investigates. The Preparer coordinates. The coding agent implements explicitly authorized changes. I review the evidence, run tests, and make the final decision.
🧪 4. Tested it against real engineering scenarios
I created two deliberately buggy Python APIs:
- Checkout API: validation defects involving quantities, prices, and discounts.
- Inventory Reservation API: negative quantities, oversubscription, invalid stock values, and unknown SKUs.
The initial tests exposed the defects. After investigation, authorized remediation, and regression testing, the final results were:
✅ Checkout: 7/7 tests passed
✅ Inventory: 9/9 tests passed
✅ Combined suite: 16/16 tests passed
🛡️ 5. Kept humans in control
One of my biggest design decisions was to make the workflow evidence-first and approval-driven.
Agents can investigate and recommend. Code changes require explicit authorization. Test results are verified locally. The final release decision stays with the human engineer.
What I learned
Switch helped me explore something more interesting than simply prompting an AI model: coordinating multiple specialized agents around a shared engineering objective.
The biggest takeaway? The real power of AI agents isn't just what one agent can do. It's what specialized agents can accomplish together when their responsibilities, communication, and evidence are clearly defined.
That's how I switched from AI conversations to AI collaboration.
Built with ❤️ using Switch.
Top comments (0)